Real-time deep learning-based vascular recognition in clinical breast photoacoustic imaging.

Breast imaging is a primary application of photoacoustic computed tomography. However, in current clinical environments, existing technologies cannot perform real-time vascular recognition during the scanning process. We propose a deep learning-based vascular recognition method. A physics-informed simulation workflow was constructed to train the network, alleviating the challenge of scarce high-quality clinical data. By optimizing the proposed HAR-UNet (Hybrid Attention Residual U-Net), an end-to-end processing speed of 18.08 fps was achieved. This high-frame-rate output enables real-time clinical vascular recognition. Furthermore, the clinical translational potential of this method was evaluated on an in vivo cohort of breast cancer patients. The proposed framework enables synchronous clinical vascular scanning and recognition. Quantitative analysis performed on 23 valid cases effectively assessed tumor angiogenesis biomarkers, differentiating tumors from normal tissues with a P -value of less than 0.0001. These results suggest that this method can provide real-time imaging assistance for the clinical evaluation and diagnosis of breast cancer.